Download Open access

A Robotic Eye Controller Based on Cooperative Neural Agents
Oscar Chang, Pascual Campoy Member, IEEE
Carol Martı́nez Student Member, IEEE and Miguel A. Olivares-Méndez Student Member, IEEE
Abstract—A neural behavior initiating agent (BIA) is proposed to integrate relevant compressed image information
coming from others cooperating and specialized neural agents.
Using this arrangement the problem of tracking and recognizing
a moving icon has been solved by partitioning it into three
simpler and separated tasks. Neural modules associated to
those tasks proved to be easier to train and show a good
general performance. The obtained neural controller can handle
spurious images and solve an acute image related task in a
dynamical environment. Under prolonged dead-lock conditions
the controller shows traces of genuine spontaneity. The overall
performance has been tested using a pan and tilt camera
platform and real images taken from several objects, showing
the good tracking results discussed in the paper.
keywords : Computer vision, neural agents, behavior initiators, agents cooperation, genuine spontaneity,neural networks.
to model and put to use these phenomena have been reported.
In [9] fuzzy rules are used to infer in the behavior of a navigation expert system. In [10] the behavior-release mechanism of
the basal ganglia was used in a robot controller. In [11] it was
experimentally demonstrate that behavior initiating agents
based upon ANN and working in collaboration with other
image processing agents, produce highly proactive vision
guided robots.
For this paper the quality of being and agent implies the
capacity to perform a useful job without external intervention
and satisfying the four weak conditions of [12]: Autonomy,
Social ability, Reactivity and Pro-activeness. In isolated
conditions some of our agents have zero reactivity or zero
proactiveness and can be described as neural modules.
I. I NTRODUCTION
II. B EHAVIOR I NITIATING AGENTS
Nature invented an acute mobile eye and a big spark toward intelligence flared in the evolutive night. Indeed experts
believe [1] [2] that the ability to carry out survival related
task in a dynamical environment, with the help of an acute
vision system, is a fundamental issue in the development
of true intelligence Sharp vision depends on complementary
image handling abilities such as efficient tracking and fast
recognition of predator/prey situations. In any case the associated controller has to respond with a quick, conflict free
behavior to images provided by the eye. In many situations
the objective moves so fast that the eye correlated muscles
are unable to mechanically compensate the displacement and
as a result the controlling vision system has to deal with an
image that travels through the associated retina. Biological
vision systems have evolved diverse mechanisms to handle
image movement and recognition. Modeling and application
of endocrine principles been reported in [3]. Special purpose
algorithms for fast pixel tracking have been studied in [4]. In
[5] competitive artificial neural networks (ANN) are devoted
to eye tracking in video sequences and in [6] a convolutional
neural network is trained for tracking purposes.
Biological neural controllers conceals a rich repertory of
behavior initiating agents which make real life neurons tick
with enthusiastic self determination. At the highest level
of brain development the existence of specific behavior
initiating mechanisms in monkeys have been studied and
modeled [7]. Recent experiments have found clear indication
of initiating mechanisms in the Drosophila brain [8]. Efforts
In the drosophila brain studies [8] researches found non
linearities and the possible competence of only a few neurons
in the final behavior initiating mechanism, deep buried in
the flys brain. Such mechanism or agent provide the fly with
genuine spontaneity -a distinctive label of living creaturesenabling the insect to get bored about tedious situations
and deciding (self decision) to take radical (and possibly
life saving) changes in its current line of behavior. These
biological behavior initiators mechanisms also produce, in
the long run, a conflict-free time sequence of behaviors,
which preserves the insect from physical damage.
The first concern of this work is to define a behavior
initiation mechanism which , mimics the capacity of genuine
spontaneity found in flies, and use it to auto impulse robotic
eye activities. The presented solution is based upon an n-flop,
a robust neural network constructed with sigmoidal neurons
and sharing a common excitatory input K . Being robust,
it serves as foundation for other large scale optimization
structures such as the TSP neural solver. The n-flop is the
basic building block beyond the concept of programming
with neurons [13], the term is derived from flip-flop a
computer circuit that has only two stable states. n-flops have
n stable states and the capacity to solved high dimentional
problems [14].
In an n-flop, neurons are programmed by their interconnections to solve the constraint that only one of the
n will be active when the system is in equilibrium. To
this end, the output of each neuron is connected with an
inhibitory input weight (-1) to each of the other n- 1 neurons
inputs (lateral inhibition). In addition each neuron receives a
common excitatory input K which, on controlled situations,
tends to force all neuron outputs towards 1. A solution is
All the authors are with the Computer Vision Group, Universidad Politecnica de Madrid, ETSII, Madrid - 28006, Spain; email:
[email protected]
978-1-4244-8126-2/10/$26.00 ©2010 IEEE
found by rising K and forcing all neurons to some nearequilibrium but unstable ”high-energy” state. At this point K
is set to almost cero, releasing the network to seek a lowenergy stable or equilibrium state. The solution given by a
non biased n-flop is an unique but unpredictable winner. A
unique winner guarantees a conflict-free operation in terms
of robotic actuators. In our terminology an isolated n-flop
behaves as an independent agent with high proactiveness and
null reactiveness, capable of producing random sequences
of conflict-free states. Proper biasing can launch statistical
reactiveness in this primal behavior.
The n-flop accepts elaborate data structured modulation
which makes possible to solve resource allocation problems
such as the TSP [13]. When properly done modulation
changes the statistics of the n-flop behavior from pure
random to quasi optimal. In terms of agent operations this is
equivalent to go from pure proactive to reactive-proactive
behavior. We use the n-flop as an integrating, modulable
behavior initiating agent, modulated with compressed image
information coming from others agents. In previous experiments this arrangement showed encouraging results [11].
Behavior initiating agent
Image centering agent
• Image recognition agent
The behavior initiating agent or BIA is based upon an
n-flop and acts as the main control administrator. In non
modulated conditions this agent produces a conflict-free,
random eye movements by its own.
The image centering agent or ICA is a trained ANN
specialized in processing the image information that conveys
the relative position of an arbitrary image in the retina. Its
job is to help the main controller to bring interesting images
to the center of the retina.
The image recognition agent or IRA is a trained ANN specialized in recognizing a particular kind of object (icon) and
discriminates it from other objects or from the background
noise. Its job is to help the main controller in the recognition
of proper icons.
Once all participant agents are debugged and tested structural modulation is activated so that the behavior initiating
agent is modulated with information coming from image
processing agents (figure 1).
•
•
A. Image Compressing Agents
For this paper an image compressing agent is a trained
neural network with the following characteristics:
• Uses a 100x 100 pixels input retina
• Has one hidden layer
• Has a small number of output neurons (less than twenty
for our current modeling)
This finite network has to be successfully trained with
backpropagation to execute a task which requires image
processing capabilities [15].
We will use the term image compression to denote that
the high dimensional set of all possible definable images
is mapped into the outputs of a classifier network with a
few output neurons. In practice all high dimensional image
activities are compressed into the conduct of a few individuals. In this work the main role of an image compressing
agent is to be trained in object recognition or object tracking
duties, in order to produce compressed image information
used, thereafter, to modulate a behavior initiating agent.
III. M ETHOD
Our goal is to construct neural machinery capable of
controlling a robotic eye that executes basic image oriented
duties while maintaining genuine spontaneity. The system
works independently, with no intervention from higher level
layers, although due to the neural nature of the controller
such intervention is feasible. For now the controller should
actively search for a particular kind of moving object and do
tracking and visual servoing of it for as long as possible.
Is the object is lost the eyed should assume a proactive
search. The system has to be able to escape from unavoidable
mistakes and traps that occur in image recognition based
upon trained ANN. The control activities are partitioned into
three neural sub-agents:
Fig. 1. Schematic of the robotic eye controller composed by three cooperating agents and a 100x100 retina, which provides collective information
A. The Behavior Initiating Agent
The image capturing system is constructed with a camera,
mounted in a pan and tilt platform (figure 2). The servo
motors are connected through non intelligent buffers to the
output of two independent 5-flops fired in parallel. These
flops acts as a behavior initiators agent (BIA) and the output
(winner) of each 5-flop is used to move the camera in
the x,y axis (horizontal/vertical) according to the 5x5 grid
arrangement shown in figure 2. Here after a given firing the
neurons x1 , y3 have become winners. Non intelligent buffers
realized plain calculations and move the servos in order to
bring this virtual point to the center of the retina. When flops
are fired again a new virtual point in the grid appears, buffers
make calculations and servos respond accordingly. When the
BIA is repeatedly fired the system respond with a random
eye movement that chases an virtual moving point. End-ofrun
information which come from the buffers is fed back toward
the n-flops so that in free running conditions the eye improves
its random search (figure 2). Considerations about this search
are given in [11].
Fig. 2. Behavior initiating Agent. The output of two isolated 5-flops is
used to activate two servomotors according to a pre established 5x5 grid
arrangement. When flops are repeatedly fired the servomotors begins to
chase and imaginary moving point.
B. The Image Centering Agent
It is well known the importance of image centering in
image recognition processes [16]. Based in our experimental
results ANN are much more difficult to train, by any method
when the processed images move too much in the retinal
area. So a training scheme were the image to be recognized
stays around the center of the retina is adopted. This raises
the recognition rate dramatically but also makes necessary to
pull any interesting image to the retina center before it can
be recognize.
Image centering is done by providing compressed image
information to the previously defined BIA, which remains
the main administrator of camera movement. Compressed
image information comes from an ANN trained with a copy
of the grid system used in by the behavior initiating agent.
A figure (circle) of variable size is randomly located in a
i,j grid position and the respective information is presented
as targets to a trainable neural net with 10 outputs, which
map the 5x5 grid positions. In the example of figure 3 a
circle has been located in the x1 ,y3 position. From this the
10 required targets are generated. This net has 19 hidden
and 10000 inputs neurons. Training is done with standard
backpropagation in bout 100.000 cycles. After training the
output neurons indicate the grid position of a received image.
The output becomes noisy and contradictory when complex
images circulate in the retina. Figure 3 shows schematics.
Fig. 3. The image Centering Agent ICA. An ANN is trained to indicate
image location in a 5x5 grid superimposed over the 100x100 retina. For
isolated, not too big images obtained information is reliable. For complex
imagery, however, information becomes very noisy and contradictory.
of different size, are presented as counterexample for arrows.
Training is done with standard backpropagation and it takes
about 100.000 cycles to get a properly qualified individual.
After training, the network successfully recognizes fat arrows
located near the center of the retina in any position. It also
clearly rejects other icons such as circles or rectangles of
any size. Although trained as a classifier the output become
noisy and falls in contradictions when complex images reach
the retina. Figure 3 shows the schematic of a qualified image
recognition agent.
C. The Image Recognition Agent
The image recognition agent is defined by the 100x100
retina input, 19 hidden and two output neurons. Its duty is
to recognize a specific icon or object. For control applications
we choose this icon to be a fat arrow. When training, a figure
of variable size and randomly rotated, is randomly located
near the center of the retina. If the figure is a fat arrow the two
targets are fixed at 0.9–0.1. Otherwise targets are fixed at 0.1–
0.9. A total of 100 objects, including circles and rectangles
Fig. 4. The image recognition agent IRA. An ANN is trained to recognize
a unique icon (arrow) against other 100 icons For isolated images near
the center of the retina the obtained information is reliable. For complex
imagery, information becomes very noisy and contradictory.
D. Agents Cooperation: The Modulation Process
In an isolated n-flop the probability of neuron i to become
a winner depends on its internal potential V(i), given by :
V (i) =
n
aj ∗ (−1) + K ∀ j = i
(1)
j=1
where:
• V(i) is the internal potential,
• aj the output from neuron j
• K common excitatory input
For modulation purposes to this internal potential a modulating scalar bk is added:
V (i) =
n
aj ∗ (−1) + K + bk
(2)
j=1
The scalar bk is in turn the weighted summation of signals
coming from the image centering agent.
That is:
10
bk =
oj ∗ wj
(3)
where
g=average of error
• o1j =output 1 of IRA
• o2j =output 2 of IRA
this average is compared against an operative threshold
which indicates how likely the icon analyzed in the last ten
BIA firing a correct one, i.e.
•
if g > threshold then close switch
(6)
else open
The appropiate setting of inter agent weights and operative
threshold causes that when the centralized icon is the correct
one the BIA accepts centering information and performs
real tracking. If observed icon is not correct, centering
information is disconnected and BIA initiates virtual eye
movements (improved random search). See figure 5.
j=1
where
• bk modulating scalar
• oj jth output of the ICA
• wj inter agent connecting weight
Equation 3 allows that the compressed image information,
coming from the ICA, modifies the statistical behavior of the
BIA.
The selection of appropriate inter agents weights wj is
critical. They change the behavior of the BIA from being
pure proactive and random, to being reactive-proactive or
only reactive. In living creatures a good balance between
these two characteristics is carefully tuned by evolution [8].
In the previous work, the selection of the inter agent
weights was done with simulated evolution [11]. For this
paper reverse engineering is used and the robotics eye is
manually tuned until it reaches its top working capacity.
Further improvement using simulated evolution might be
considered.
In order to simplify the image centering modulation in
equation 3, just the weights bettwen conected alike neurons
in the ICA and the BIA, are set different from cero:
bi = oj ∗ wi
only for
i=j
(4)
This leaves only ten weights which were experimentally
set to 0.1.
Image recognizing modulation:
To complete te modulation process each inter agent weight
wj is in turn modulated (switched on-off) by the average
value of the difference of the image recognizing agent in the
last ten BIA firing that is, let
0
1 g=
o1j − o2j
10 j=−10
(5)
Fig. 5. Modulated Robotic eye controller. BIA accepts compressed image
information from cooperating ICA. This information is enough to carry on
an effective image tracking. The image recognizing agent (IRA) controls
through average error is an found icon deserves to be tracked or not. When
not tracking BIA performs imaginary eye movement.
IV. R ESULTS
The system behaves well with simulated and real images
and shows some promising capacities. For instance when the
eye explores a complex background formed by superimposed
icons, some spurious figures are mistaken as arrows, causing a potential deadlock. In this condition the controllers
mimics a living drosophila and from time to time changes
spontaneously its behavior trying to find a way out (or so it
seems to an external observer). Under complex imagery lost
and recuperation of icon frequently occur. In such cases the
controller keeps a hard working, proactive attitude and, in
our experiments, no continuous lost of the icon ever occurs.
In figure 6 a typical tracking sequence is shown. The image
recognition agent has been deactivated and the controller
has to track-servoing a moving circle which changes in size
and have some associate noise. During the tracking the BIA
Fig. 6. The robotic eye controller realizes a tracking-servoing exercise by
following a moving circle which changes in size and has some lateral noise.
The two servo positions are generated by the BIA with the help of centering
information coming from the ICA.
generates the x-y servos positions with the help of centering
information coming from the ICA.
Figure 7 shows the x-y servo activities with the three
participating agents fully activated. Icons move up to 10
cm/seg and rotate up to 60 rpm., images are processed
at about 20 frames/sec. When and arrow is spotted servo
behavior is modulated by centering information and tracking
begins. If the arrow is lost servos change their behavior to a
semi-random search which produces many spurious images.
separated tasks that can be executed by three independent
neural agents. Neural agents associated to those simpler
tasks proved to be easier to train and show a good general
performance. The proposed behavior initiating agent (BIA)
provides and integrating low bandwidth channel where others
specialized agents can contribute with relevant compressed
information. As a result the obtained controller accomplishes
an acute image related task, in a dynamical environment
without requiring the intervention of higher control layers. As
a disadvantage each decision taken by the controlling neural
machinery remains stochastic variable so erroneous decisions
will always be present. This behavior is shared with living
organisms. On the other hand the BIA has the capacity to
successfully keep on moving the robotic eye under very noisy
conditions and for long periods. Under such circumstances
spurious images are usually deadly traps for image-driven
machinery. This enduring capacity appears because under
appropriate modulation conditions the BIA keeps a dash of
genuine spontaneity.
ACKNOWLEDGMENT
The authors would like to thank the Spanish Science and
Technology Ministry for funding the experiments of this
work through the R&D project CICYT DPI 2007-66156. The
authors also like to thank UPM for the grant as Visitant
Professor of the first author and also the Consejera de
Educacin del a Comunidad de Madrid and the Fondo Social
Europeo (FSE) for some of the authors PhD Scholarship.
R EFERENCES
Fig. 7. Typical search-track sequence completed by the proposed neural
machinery. When an arrow is spotted the BIA accomplish tracking. If the
arrow is lost the BIA adopts a semi-random search. Many spurious images
are found and processed during the sequence. Icons move up to 10 cm/seg
and rotate up to 60 rpm.
V. C ONCLUSION
The problem of tracking and recognizing a moving target
has been solved by partitioning it in three simpler and
[1] R.
A.
Brooks,
“Intelligence
without
representation,”
1991,
no.
47,
pp.
139–159.
[Online].
Available:
http://citeseer.ist.psu.edu/old/brooks91intelligence.html
[2] H. P. Moravec, “Locomotion, Vision and Intelligence,” in Robotics
Research-,The First International Symposium, M. Brady and R. Paul,
Eds. Cambridge, Ma: MIT Press, 1984, pp. 215–224.
[3] J. J. Henley and D. P. Barnes, “An artificial neuro-endocrine kinematics
model for legged robot obstacle negotiation.” 8th ESA Workshop
on Advanced Space Technologies for Robotics and Automation, Nov
2004.
[4] F. Dellaert and R. Collins, “Fast image-based tracking by selective
pixel integration,” in ICCV 99 Workshop on Frame-Rate Vision,
September 1999.
[5] M. G. Marcone, G. and L. L., “Eye tracking in image sequences by
competitive neural networks,” Neural Processing Letters, vol. 7, no. 3,
pp. 133–138, 1998.
[6] D. D. Lee and H. S. Seung, “A neural network based head tracking
system,” in NIPS ’97: Proceedings of the 1997 conference on Advances
in neural information processing systems 10. Cambridge, MA, USA:
MIT Press, 1998, pp. 908–914.
[7] L. D. Soltani, A. and W. Xiao-Jing, “Neural mechanism for stochastic
behavior during a competitive game,” Sciences Direct, Neural Networks, vol. 19, no. 8, 2006.
[8] A. Maye, H. Ch., G. Sugihara, and B. B., “Order in spontaneous
behavior,” PLoS One, vol. 10, no. 1371, 2007. [Online]. Available:
http://www.plosone.org/article/info:doi/10.1371/journal.pone.0000443
[9] G. V. Parasuraman, S. and B. Shirinzadeh, “Fuzzy decision mechanism combined with neuro-fuzzy controller for behavior based robot
navigation,” Industrial Electronics Society, 2003. IECON apos;03. The
29th Annual Conference of the IEEE, vol. 3, no. 2, 2003.
[10] T. J. Prescott, M. F., G. K., H. M., and R. P., “A robot model of
the basal ganglia: Behavior and intrinsic processing,” Sciences Direct,
Neural Networks, vol. 19, no. 1, 2006.
[11] O. Chang, “Evolving cooperative neural agents for controlling vision
guided mobile robots,” Proceedings of 2009 8th IEEE International
Conference on Cybernetic Intelligent Systems, September 9TH-10Th,
2009,Birmingham University, UK, 2009.
[12] M. Woolridge and N. R. Jennings, “Intelligent agents: Theory and
practice,” Knowledge Engineering Review, vol. 10, no. 2, 1994.
[13] J. B. Shackleford, “Neural data structures: programming with neurons
- technical,” Hewlett-Packard Journal, pp. 69–78, 1989.
[14] G. Serpen, “Managing spatio-temporal complexity in Hopfield neural
network simulations for large-scale static optimization,” Mathematics
and Computers in Simulation, pp. 279–293, 2004. [Online]. Available:
http://dx.doi.org/10.1016/j.matcom.2003.09.023
[15] D. Pomerleau, “Knowledge-based training of artificial neural networks
for autonomous robot driving,” in Robot Learning, J. Connell and
S. Mahadevan, Eds., 1993, pp. 19–43.
[16] R. J. Vogelstein, M. U., E. Culurciello, R. Etienne-Cummings, and
G. Cauwenberghs, “Spatial acuity modulation of an address-event
imager,” in Proceedings of the IEEE International Conference on
Electronics, Circuits and Systems. IEEE, 2004, pp. 207–210.